Config class and an async run function. Condor runs it directly: it fetches data, computes what you asked for and returns a result or saves a report. No LLM reads the code or interprets the output while it runs, so the same input gives the same output and a run costs no tokens.
Use a routine for work you want done the same way every time: scans, indicators, price checks, bot summaries, backtest charts. Use a skill when the work needs judgment.
Routines and skills
Where routines live
Who can run a routine depends on which folder it’s in.<condor> is the folder you installed Condor into.
When two folders hold a routine with the same name, the more specific one wins:
- An agent’s own routine shadows a shared routine with the same name.
- For the chat, a routine in
.condor/agents/condor/routines/shadows a shared or shipped one. - A routine in
.condor/agents/shadows the shipped copy inagents/.
routines/ folder. See Architecture for how agent folders work.
A minimal routine
- The file name is the routine’s name. Save this as
price_check.pyand it appears asprice_check. - The first line of the
Configdocstring is the routine’s description. It shows on the routine’s card and in the list agents see. - Each
Field(description=...)labels a setting in the Workspace config form and in Telegram. get_client(chat_id, context=context)returns a Hummingbot API client for the server the run was started on.
What run returns
run(config, context) can return either:
- A string. Shown in Telegram and in the Workspace’s Output view.
- A
RoutineResult. It has atextfield for Telegram and optional fields the Workspace renders:table_dataandtable_columns, achart_image(PNG bytes) andsections.
One-shot and continuous routines
- One-shot routines run once and return. You can run them on demand or on a schedule.
- Continuous routines set
CONTINUOUS = Trueat module level. They loop until you stop them, and they handle their own timing withasyncio.sleep. When stopped, Condor cancels the task, so catchasyncio.CancelledErrorand return a closing message. Continuous routines can’t be scheduled.
price_monitor routine is continuous. It keeps a live report up to date while it runs. Creating Routines has a full example.
Optional module attributes
Running routines
When an agent runs a loop, its prompt lists every routine it can run, with each routine’s description. A Dry Run refuses routine runs, because a routine can run any Python. Listing routines and reading their source still work.
No restart needed
Condor checks the routine folders each time it lists routines, and reloads only files whose modification time changed. A new or edited file shows up the next time you open the Routines page, open/routines in Telegram, or an agent lists its routines.
Routines inside an agent you imported, for example from the Agent Hub, stay pending until you allow them in that agent’s Routines tab. Routines you create yourself load as above.
Telegram’s 🔄 Reload button forces a full reload. Use it to retry a file that failed to load.
Let Condor write routines
You can describe a routine in the chat and Condor writes it, loads it and runs it to test it:.condor/agents/condor/routines/, which git ignores. To make one available to every agent, ask Condor to publish it as shared, which saves it to .condor/agents/_shared/routines/. An agent that writes a routine saves it to its own routines/ folder.
Next steps
Creating Routines
Write, test and use your own routines
Built-in Routines
The routines that ship with Condor
Reports
Save HTML reports with charts and tables
Routines page
Run, schedule and browse reports in the Workspace

